EP0357212B1 - Artificial neural device and corresponding method - Google Patents
Artificial neural device and corresponding method Download PDFInfo
- Publication number
- EP0357212B1 EP0357212B1 EP89307440A EP89307440A EP0357212B1 EP 0357212 B1 EP0357212 B1 EP 0357212B1 EP 89307440 A EP89307440 A EP 89307440A EP 89307440 A EP89307440 A EP 89307440A EP 0357212 B1 EP0357212 B1 EP 0357212B1
- Authority
- EP
- European Patent Office
- Prior art keywords
- stimulus
- response
- signal
- patterns
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Lifetime
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04R—LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; DEAF-AID SETS; PUBLIC ADDRESS SYSTEMS
- H04R25/00—Deaf-aid sets, i.e. electro-acoustic or electro-mechanical hearing aids; Electric tinnitus maskers providing an auditory perception
- H04R25/50—Customised settings for obtaining desired overall acoustical characteristics
- H04R25/505—Customised settings for obtaining desired overall acoustical characteristics using digital signal processing
- H04R25/507—Customised settings for obtaining desired overall acoustical characteristics using digital signal processing implemented by neural network or fuzzy logic
Definitions
- This invention relates to artificial neural devices and in particular relates to artificial neural devices having enhanced storage and processing capabilities.
- Artificial neural systems embody a form of artificial intelligence which attempts to emulate the way in which biological neurons or group of neurons process information.
- Artificial neural systems may be embodied in computational hardware and/or software which are capable of processing and storing information.
- ANS devices can store a relatively large number of patterns, so as to quickly produce a particular response when the ANS device is stimulated by a particular stimulus input.
- ANS devices immediately map input patterns into the nearest stored pattern.
- EP-A-0 085 545 discloses a pattern recognition apparatus including a vector generating unit for generating an input vector representing the characteristics of an unknown input pattern, a dictionary unit which stores a plurality of reference vectors for each category, a similarity calculating unit which calculates a similarity between the input vector and a plurality of reference vectors for each category, a similarity calculating unit, and an additional dictionary generating unit which generates additional reference vectors for a particular need or application.
- Document US-A-4 366 551 describes a knowledge storage and retrieval method performed to result in a single address number of a storage region which represents a large quantity of information.
- Document US-A-4 523 331 discloses a computer algorithm and apparatus for automated image input, storage and output. Each image is transformed into a unique binary number and then stored as such. This unique binary number is obtained by adding unique binary value of each pixel present in that image. Means for processing handwriting and coloured images are also disclosed. Image recognition and matching takes place by comparing the binary value of the new image received against all images stored in the descending order of difference in binary values.
- the device described herein employs a novel implementation for storing and processing information. Advantages and operational characteristics realized by at least preferred embodiments of the invention described herein are outlined below:
- One aspect of this invention relates to an artificial neural device comprising and method as set out in claims 1 and 6 respectively.
- the artificial neural system described herein may be applied for innumerable applications including:
- Table 1 represents the probalistic standard deviation error within the response as a function of the number of encoded patterns.
- Table 2 list of phoneme codes.
- Figure 1 illustrates in general the components of the biological neuron.
- Figure 2 illustrates in general the sigmoid response characteristics and signal processing characteristics of the biological neuron.
- Figure 3 illustrates an embodiment of hardware configuration for the artificial neural device illustrative of a single neural element.
- Figure 4 illustrates in general an example of a hardware configuration for a visual to auditory artificial neural device.
- Figure 5 illustrates a digitized visual image represented by pixels within the visual field.
- Figure 6 illustrates a digitized sonogram for auditory input.
- Figure 7 illustrates the application of an artificial neural system within an automatically controlled transportation vehicle.
- Figure 8 illustrates patterns that correspond to incremental movements of a robotic system over a given time frame.
- Figure 9 illustrates a decision tree
- Figure 10 illustrates a multilayered network
- Figure 11 illustrates a layered parallel network.
- Figure 12 illustrates a feedback network
- Figure 13 illustrates an embodiment of hardware configuration for the artificial neural device for a single processing node.
- Figure 14 illustrates an embodiment of a 16 node hypercube configuration for link communication between processing nodes.
- a radically new approach to ANS theory is described herein.
- a general conceptual description is provided followed by a mathematical description.
- a principle goal in the field of ANS research is the development of a mechanism to encode or superimpose a large amount of information onto a given data storage medium and facilitate a means to decode that stimulus input, via the encoded information, to produce a response output.
- An optical hologram provides a system roughly analogous to the above.
- the information encoding mechanism within the optical hologram permits a visual 3-dimensional object to be stored onto, and reconstructed from a 2-dimensional photographic plane effectively enfolding or collapsing an apparent dimension.
- information storage densities within the 2-dimensional plane are greatly enhanced.
- stimulus-response patterns may represent any form of sensory modality and may be read-in either via external environmental receptors or (in the case for multineuronal systems) propagated through the system via analogous interneuronal synaptic connections.
- the process superimposes these stimulus-response patterns onto the same memory or storage medium.
- the theoretical basis is roughly analogous to the optical hologram, with a principal exception that the time index is the apparent dimension enfolded or collapsed.
- the stimulus On activation of the system by a stimulus pattern (decoding operation) the stimulus is processed via the accumulated information (all encoded stimulus-response patterns) to issue a near instantaneous response.
- the characteristics of the response output are described in the section pertaining to Decoding.
- patterns refers to sets of data and the term will be used to indicate same.
- the data values contained within both the stimulus and the response patterns may represent any form or type of information.
- the sets of data within the stimulus or response patterns may represent pixel values within a visual field, axis position or rates of movement within a robotic device, or Fourier transform elements for auditory input.
- both the stimulus and the response field may contain data representing any form of sensory modality, measurable parameter, or abstract representation of the above.
- the stimulus signal or field will be represented by the following data set [S]' and will be represented as a 1 by N matrix.
- matrix [S]' may represent the numerical data produced by a video digitizing equipment when viewing an object such as an apple.
- [S]' [s 1 , s 2 , s 3 , ...s N ] where s 1 to s N is a set of numeric or scalar values.
- such numerical values can represent a character representation of the word APPLE.
- the process involves representing the information contained within both of the above sets, namely the stimulus signal or field [S]' and response signal or field [R]', as phase angles corresponding to vector orientation on a two dimensional or complex (Argand) plane.
- This may be performed by translating the above sets of data to the following vector matrices (1 by N, or M) whereby s N and r M values are normalized to, or scaled within a circular range corresponding to 2 ⁇ or 360°, thus producing a proportional phase angle representation as denoted below by ⁇ N and ⁇ N respectively:
- [S] [e i ⁇ 1 , e i ⁇ 2 , e i ⁇ 3 , ... e i ⁇ N ]
- [R] [e i ⁇ 1 , e i ⁇ 2 , e i ⁇ 3 , ... e i ⁇ M ]
- the above complex exponential function can be represented in several other forms (i.e. Z, 1nZ, a + ib, etc.) however the exponential form has been used for convenience.
- Magnitude components may be associated with each element or vector within the stimulus [S] and response [R] field, however for simplification of this illustration, and the technique in general, the vector magnitudes are assumed equivalent to unity.
- Encoding of the network is performed by determining the difference between the vector orientation or phase angle between every element (or subset of elements) in the stimulus input field [S] and every element (or a subset of elements) in the output response field [R].
- transpose of [A] is obtained by interchanging the rows and columns of [A].
- the multiplication of two matrices [A], [B] may be performed where the number of columns of the first matrix [A] is the same as the number of rows in the second matrix [B].
- the element in the jth row and the kth column of [M] is determined by multiplying each element of the jth row of T / [S] by the corresponding element of kth column of [R] and then summing the resulting product terms. This operation may be illustrated for one pattern as follows: where
- Re(m j,k ) Re( s 1,j )xRe(r 1,k )-Im( s 1,j )xIm(r 1,k )
- Im(m j,k ) Re( s 1,j ) x Im(r 1,k ) + Im( s 1,j ) x Re(r 1,k )
- m j,k ⁇ j,k e i ⁇ * jk
- ⁇ j,k [Re(m j,k ) 2 + Im(m j,k ) 2 ] 1/2
- [M] represent the encoded information for one (1) stimulus-response pattern.
- [M1] shall denote the correlation matrix for this pattern 1.
- the above technique may be expanded to encode a large number of stimulus-response patterns onto the same memory space.
- (x j,k ) sums the vector differences between element j in the stimulus field and element k in the response field over all encoded patterns.
- the decoding operation is related to the associative recall of the appropriate encoded on "learned" response.
- This operation consists of obtaining or reading data corresponding to an input stimulus. In other words one may focus the video digitizing equipment upon the image of an apple to generate the associated numerical data representative of the visual image.
- Equation 19 may be expressed in the following summation form: again where:
- each element of the response vector matrix [R] may be determined from the following equation:
- the phase angle within the generated output response will be equivalent to the encoded or learned response phase angle for pattern p* including, however, a deterministic error contribution.
- This error contribution increases with the number of patterns stored and the vector summation properties as illustrated by equation 21 are roughly analogous to the characteristics of Brownian motion for the uncorrelated patterns.
- the deterministic error contribution as a function of the number of elements within the stimulus field (N), the number of patterns stored (P), and for P/N ratios less than .25, may be approximated by:
- the encoded stimulus-response patterns may represent completely non-linear and independent functions.
- the error term within the output response is deterministic.
- the characteristics at this "error" term and a general descriptive form for the decoded response elements (r k ) is illustrated in a more rigorous fashion below.
- pattern variance term will be defined as the degree of correspondence between a stimulus pattern encoded within matrix [X] and the input stimulus pattern [S]*.
- This pattern variance may be defined mathematically as follows:
- Each of the summation terms in equation 24 displays a varying dominance or magnitude ( ⁇ p ) which statistically is inversely proportional to the pattern variance (eq. 23).
- ⁇ p the encoded patterns most similar to the input stimulus pattern [S]* produce the more dominant contribution within the generated response output (r).
- N / C gives the magnitude component for the pattern (p*) within equation 24 and represents the most dominant vector for a perfectly correlated pattern match.
- the phase angle ⁇ p * corresponds to the encoded response ( ⁇ p *) for pattern p *.
- the generated response phase angle ⁇ p * is deterministic but largely independent or uncorrelated with respect to the encoded response ( ⁇ p *).
- stimulus-response patterns over a series of frames (generally with respect to time), one may represent the stimulus field as a two dimensional complex vector matrix [S] with the set of vectors (complex numbers i.e. e i ⁇ 1 to N ) representing each element of the stimulus field along the row or horizontal, and the pattern index or "time" along the column or vertical i.e.:
- a VECTOR MATRIX shall be defined as a matrix whose elements are vectors (i.e. complex numbers having a particular orientation on the Argand plane).
- response field [R] corresponding to the above stimulus patterns could be represented in a similar fashion but, for purposes of illustration, will be assumed to have only a single element. Again the pattern or time index is along the column or vertical, i.e.:
- the method by which the process operates is that a response value corresponding to any of the stored stimulus-response patterns can be regenerated at any time by inputting a previously encoded stimulus pattern to the device (again recalling that the stimulus pattern could be a picture, a spoken word, or other parameter, and the response could be the associated word, control action, respectively).
- This response decoding operation is performed as represented by equation 19 and contains an associated deterministic error as approximated by equation 22.
- each neuron is bounded by a cell membrane and contains a nucleus.
- the size and the shape of these neurons vary widely, but the structural plan always contains certain elements as shown in Figure 1: a cell body or soma, and the processes from this cell body, namely an axon (neurite) and usually several dendrites.
- the classification of the neuronal processes in terms of an axon and several dendrites is made on the basis of function.
- the axon links the nerve cell with other cells.
- the axons of other cells terminate on the dendrites and also on the soma.
- the axon and dendrites normally divide into a varying number of branches after emerging from the soma.
- the junction of an axonal ending with another cell is called a synapse.
- Neurons may have up to several thousand synaptic connections depending upon function.
- Information propagation within the neuron occurs in the following general fashion; with signals relayed in the form of pulse modulated waveforms received via the dendrites and soma and mediated via the synaptic connections. Conversely, in response to the input signals received as above, the neuron generates an output pulse modulated waveform which travels via the axon to subsequent layers of neurons.
- the general structure of the neuron with respect to the signal pathways is illustrated in Figure 2.
- the biological neuron stores information in the form of stimulus-response patterns and follows a similar encoding/decoding mechanism as defined in the section pertaining to theory.
- each synaptic connection within the dendritic structure of the neuron receives a pulse modulated wave form characteristic of Figure 2.
- These waveforms may be represented in a general and abstract form with pulse frequency representative of a vector phase angle and pulse magnitude representative of a vector magnitude.
- [X] contains the encoded set of stimulus-response patterns, represented in an abstract form (complex numbers or vectors). The corresponding information is superimposed upon the same storage medium, that is, the synapse or area of the synaptic connection.
- each neuron may operate in an independent or non-syncronous manner.
- the interneuronal synaptic connections may be configured to produce a highly associative memory capability, extending across all forms of sensory modalities.
- the above process defines a mechanism for storage for an extremely vast amount of information, potentially analogous to a lifetime of accumulated stimulus data, again derived from all forms of sensory modalities.
- the decoding process illustrates a mechanism by which the neurobiological system could process incoming sensory stimulus through the above accumulated lifetime of encoded information, to issue a response in a near instantaneous manner.
- the hardware embodiment is illustrated in Figure 3. In this example, sixteen stimulus inputs are shown. It is expected that analogue signal multiplexing techniques would be employed to increase the number of elements within the stimulus field [S] to any desired number.
- the input signals are oscillating wave forms with the vector phase information represented by frequency and vector amplitude represented by signal amplitude. This illustration is analogous to the neurobiological waveform signal.
- the input stimulus field elements (1 to 16) are transformed into a digital representation of phase angle and magnitude via the demodulating circuit (A).
- a separate phase and magnitude component is determined for each of the input stimulus elements.
- the associated output response for encoding is read in via demodulating circuit (E) and similarly converted into a digital representation of the phase angle and magnitude components.
- the digital encoding unit (F) performs the complex vector transformation as defined within equation 35A to encode the stimulus-response pattern into storage unit (C).
- Unit (C) stores the encoded information and represents the axo-dendritic synaptic analog.
- the input stimulus signal elements are transformed into a digital representation of phase angle and magnitude via the demodulating circuit (A).
- the decoding unit (B) reads the binary representation of the input stimulus signals and reads the corresponding encoded vector elements [X] from unit (C).
- the appropriate elements of [X] may be determined via an address enable IN bus controlled by external logic.
- Unit (B) performs the decoding operation defined by equation 19 using the input stimulus elements [S] from unit (A) and the encoded elements [X] from unit (C).
- the output response is transmitted to the frequency modulator unit (D) which converts the binary representation of the vector or complex number into a waveform with frequency representative of vector phase angle and signal amplitude representative of the vector magnitude.
- neural elements may be connected together in any configuration to form an array of neural elements operating in a simultaneous or massively parallel fashion.
- the following illustrates six ANS devices that may be derived from the process referred to earlier. A virtually unlimited number of devices and applications may be conceived, however the purpose here is to illustrate the range of applications to which the ANS device as described herein may be employed.
- a vision recognition device embodying the invention as described above is schematically depicted in Figure 4.
- the visual recognition device comprises conventional hardware for capturing and input of the visual data for encoding/decoding via the system.
- This device 1 employs the use of a single transformation group (TG) to perform the visual to auditory conversion.
- TG single transformation group
- the hardware requirements for the device 1 are as follows:
- a series of phonemes comprising the output response will be encoded and the hardware handles the translation from phoneme to spoken output.
- Each spoken word consists of a list of phonemes. Approximately 36 phonemes are required to verbalize the majority of phrases within the English language.
- the visual data stream generated by the video signal digitizer 3 (series of pixel values) is translated to a verbal output (list of phonemes) at speaker 12.
- the video digitizing equipment 2 generates for example a series of 4096 bytes which represents a field 64 by 64 pixels for visual input.
- the pixel input values should be translated into vector phase angles 5 such that they fall within the range of 0 to 2 ⁇ .
- the visual input field shall be represented as [S].
- the corresponding output response 10 for that visual stimulus is either a spectral harmonic representation for the spoken word, or an equivalent list of phonemes each defined by a numeric representation; the later shall be illustrated for simplicity.
- This word may be reconstructed with 4 phonemes: h , o , u , s.
- the representation for these four phonemes must be encoded into Memory and ANS processor 1.
- a total of 36 phonemes (See Table 2) may be chosen to make up the sound therefore a selection mechanism must be determined.
- phase angle component with magnitude set equivalent to unity.
- the complex plane must be divided into 6 sections, each section separated by 60 degrees. For example, to encode a "0" a phase angle of ⁇ /3 must be assigned, similarly:
- the encoded output would consist of the following:
- the actual phoneme response could be input via the keyboard 13 and the conversion to a set of complex values (phase angle 0 to 2 ⁇ ) performed.
- a correlation matrix [X] is generated and stored in volatile memory 8 accessible to the ANS decoding and encoding processors 7 and 9.
- the stimulus-response patterns are encoded into the correlation matrix [X] via the ANS processor 9 performing the following operations. These steps are performed for each element (x j,k ) within the correlation matrix [X]:
- the above illustrates the encoding portion of the process. This process may be repeated to encode any number of stimulus/response patterns into the network.
- the ANS processor 9 should optimally be configured to perform the conversions (step 3 to 5) in as parallel a fashion as possible. A matrix processing device or configuration within the ANS processor is therefore desirable. The above operation may of course be performed sequentially as in conventional general purpose computing devices, however, processing time may become an inhibiting factor.
- a similar operation of encoding is performed on decoding a stimulus input into a response.
- a video signal is generated by video camera 2 and feeds input to video signal digitizer 3 which digitizes the image into a 64 by 64 pixel array. Again the pixel values are proportional to the gray scale or intensity level. Similarly, these pixel values are converted into vectors 5 (i.e. represented by phase angles between 0 and 2 ⁇ ).
- the stimulus field shall be represented as [S]* and elements of [S]* representing phase angles are read into volatile memory accessible to the ANS decoding processor 7. The following operations are performed on the complex elements within the correlation matrix [X] and the elements within the stimulus field [S]*.
- the above illustrates a very simple application of the ANS device employing only a single transformation group.
- This application transforms the input sensory modality of vision into the output modality of speech.
- These visual patterns may be of any form or degree of complexity (i.e. face, general scene, etc.). Due to the non-ideal distribution generated from visual objects the stimulus-response storage capacity, based on a deterministic error criterion, is somewhat reduced over the idealized case as for randomly generated patterns.
- a further embodiment related to the memory storage capacity may be performed by preprocessing both the input stimulus field [S] on encoding (step 2) and preprocessing [S]* on decoding (Step 2) into a matrix of Hermittian form.
- the resultant effect of the above Hermittian operation is that the visaul stimulus-response pattern storage capacity could exceed 1 Million prior to onset of excessive misclassification error.
- the above operations would optimally be performed by dedicated matrix processing hardware within the ANS processor device.
- This ANS device may potentially display significant advantages over current recognition schemes.
- the most significant advantage is the high pattern storage capacity.
- This system achieves a state whereby the information corresponding to several images is encoded in an abstract (complex vector) representation onto the same data storage medium [X].
- This "holographic" information storage technique presents significant advantages both in the physical data storage requirements (i.e. size of memory storage device) and the processing time requirement on decoding to a response output.
- Voice to Text Transcription is another example of the transformation of one form of input sensory modality (ie. hearing) into an alternate form (ie. visual transcription).
- This device would entail encoding and decoding of a voice input stimulus into a visual or typed output.
- the input stimulus field would consist of a digitized sonogram.
- a sonogram is one means of expressing auditory data and indicates intensities of harmonic content over a given time axis.
- An example of a sonogram is presented in Figure 6. Much like a digitized visual input field, a sonogram may be divided into a pixel field, within frequency along the vertical axis and time along the horizontal.
- the input is therefore the spoken word and the output is a typed response.
- Robotic or cybernetic devices are of course another level of complexity beyond the transformation devices described above. It is expected that for robotic or cybernetic control, arrays of transformation groups (TG) will be required, employing various feedback control modes.
- Figure 7 illustrates a one possible multi-TG configuration with feedback.
- the ANS control device as described herein can be utilized with a robotic assembly device, possessing 10 rotational axis and visual input.
- the feedback mechanism is via visual stimulus and axial joint position/rate of movement.
- This device consists of two transformation groups, TG1 and TG2.
- Transformation group would receive inputs indicating axial position (10 elements) and rates of movement (10 elements).
- This transformation group could employ Hermittian preprocessing (see section entitled Enhancements of Invention) of the stimulus field to greatly increase the pattern storage capacity.
- the input field could be expanded to 160,000 elements (ie. 20 4 ).
- This expanded stimulus field could then be supplied to the second transformation group (TG2) receiving stimulus also from a visual field (ie. 256 by 256 pixel field).
- the above model provides a total of 244,000 elements within the stimulus field of TG2 for stimulus-response encoding and decoding operations (ie. 64,000 visual and 160,000 axial position/rate).
- the method of encoding or learning dynamic control response motions for the robotic arm is illustrated simply as follows.
- the complex movement would be determined as a series of patterns representing visual input, axial position and rates of movement. These patterns would correspond to each incremental movement of the process over a given time frame illustrated rather coarsely in Figure 8.
- These stimulus pattern frames could be encoded into the network with the output response pattern corresponding to the next (time) incremental axial position or movement for the robotic device.
- the ANS device would serve as inputs the following:
- the number of input elements is expected to be large (say 200,000 elements). Again assuming an N/4 rule for this pattern distribution, the number of stimulus/response patterns that could be encoded would be in the area of 50,000. If second order Hermittian preprocessing is employed on the stimulus inputs, the pattern storage capacity could potentially exceed 1 Billion. It is expected that this storage capacity could encode sufficient control response actions to cover substantially any potential accident situation or defense manouver, and bring this application into the realm of feasibility.
- Heuristics is a term used to define the concept of rule based programming. In general, the approach involves applying multiple decision rules against input states in order to issue a seemingly intelligent response. Heuristics was initially employed within the field of game playing (ie. chess) and displayed particularly impressive results.
- Rule based programming has become the mainstream of AI research and has found more practical applications within the field of expert systems.
- a large degree of research in this field centers about the development of more efficient searching and parsing mechanisms for traversing the rule based decision tree.
- the principle drawback within the conventional heuristic approach is that decision rules must be applied and tested in a sequential fashion prior to arriving at a given response or outcome.
- the number of rules and thus search time required increases dramatically, limiting the capabilities of the rule-based approach.
- a simple, however, direct analogy between the ANS process and functional aspects of the heuristic technique shall be made.
- the general concept of the decision tree within the heuristic approach is presented below.
- the form of the decision tree is represented in Figure 9.
- the top event is defined as one possible output and the branching network below describes the boolean relationship which arrives at this top event.
- the boolean tree describes all conditioning inputs leading to the top decision event in the form of AND and OR relationships (in the simplest case). Multiple decision trees are employed within any expert system. In the example given here:
- Each of the above product terms (1 to 12) consists of a series of "anded” conditions, in other words, states that occur simultaneously in time.
- the above minimal cutsets or "input patterns” may be directly encoded into the ANS device in the form of discreet stimulus-response patterns, the response action being the top event in the decision tree.
- Each of the above product terms in fact represents a separate stimulus-response pattern and all patterns may be sequentially encoded into the ANS network.
- the ANS device described herein provides an enormous capacity for storage of stimulus-response patterns and facilitates a mechanism for construction of extremely large heuristic rule based systems.
- This ANS based expert system would therefore be capable of storing an encoded response for 250,000 scenarios, each consisting of 1000 input parameters.
- a number 2 base system shall be used for generation of the output response, therefore, 18 binary digits (log 2 250,000) are required to identify 250,000 distinct scenarios.
- the field of prosthetic control could benefit greatly from the application of the ANS device described herein. It is reiterated that the purpose and intent of this device is to provide more efficient information storage and processing capabilities via emulation of the suspected mechanism by which the biological neural system stores and processes information.
- the specific ANS device described herein appears to emulate the form of the information processing characteristics within the neuron in many respects, these similarities are outlined below:
- a synthetic system which closely mimics the above biological characteristics would inevitably provide significant advances in the area of interfacing prosthetic device control to neurobiological circuits.
- This ANS device would yield significant advances in the area of control for artificial limbs, and it is expected, would inevitably advance towards prosthetics for higher functions.
- the ANS device described herein may be modified in several areas to provide enhanced and improved operational characteristics. These enhancements may be realized in the application of Hermittian and sigmoid preprocessing methods, layered networks, modification of the magnitude components of the stimulus-response pattern elements, sparse matrix techniques, methods for establishment of a memory profile, and excitory and inhibitory synaptic connections. These are of course only a limited subset of possible enhancements and are outlined below.
- the following process vastly improves the characteristics of the ANS device particularly when the input stimulus field consists of a relatively few number of elements.
- This process greatly increases the number of stimulus-response patterns that may be stored within the artificial neural device for a given number of elements within the stimulus field.
- the process involves translating the stimulus pattern [S] into a matrix of Hermittian form.
- the result of which is the formation of a matrix of vectors or complex elements which relate every element in [S] to every other element in [S] as a difference in the vector orientations.
- [S] [ e i ⁇ 1 , e i ⁇ 2 , e i ⁇ 3 , e i ⁇ 4 , . . . ]
- the elements of the above Hermittian matrix provide the input stimulus field to the ANS device for both encoding and decoding operations.
- One example may be a system in which the two diagonals adjacent to the central diagonal are formed, to illustrate:
- FIG. 10 illustrates a layered network where TG represents a transformation group and refers to a device or set of devices that perform a simple transform operation from a single input stimulus field [S] to a single output response field [R].
- a layered network consists of a series of transformation groups TG1, TG2, TG3 in which the output response field from one transformation group or layer (TG1) feeds into the input stimulus field of a higher layer (TG2). Any number of layers may be utilized to form more complex arrays of networks.
- a parallel network as illustrated in Figure 11 may be formed in which layers within the Network consist of multiple transform groups operating in parallel.
- the output response field may feed into any of the input stimulus fields of a subsequent layer in any configuration or arrangement.
- Any configuration of both layered and parallel transformation groups may be utilized to form a highly complex array of networks.
- Figure 12 illustrates a system where the output response fields of neural elements within higher layers feed the effects back into input stimulus fields of neural elements within lower or earlier layers.
- Any configuration or arrangement of feedback systems may be employed.
- the ANS device inherently provides feedback control for any situation in which the output response manipulates or modifies the local environment providing the input stimulus, (ie. feedback via environment).
- This concept may be visualized within a robotic control systems in which the input stimulus is axial position and rate sensors and the output response is the motor control effectors. Any change in the response output (axis movement) modifies the input stimulus pattern thus effecting the subsequent output response, and so on. It is expected that the above feedback mechanisms will significantly enhance the control capabilities within most applications of the ANS device.
- a memory profile refers to a mechanism to establish a relative dominance of encoded or learned stimulus-response patterns as a function of time. Any time variable function or transfer function (ie. first order lag or decay) may be applied to any element within the correlation matrix [X] in order to establish a memory profile. An example may be illustrated assuming that the components of [X] decay with a first order exponential relationship.
- a mathematical relationship for a first order decay may be derived from equation 35B to produce a value for the encoded elements of [X] where:
- stimulus-response patterns learned at an instant one time constant (i.e. 1 hour) in the past will have a relative dominance of 0.368 in comparison to a pattern that has been encoded or learned within the current time frame (T).
- the stimulus and response information as encoded into the network corresponds to a circular range extending over 2 ⁇ on a complex plane.
- Many physical parameters are not directly amenable to scaling within a similar closed range and limitations with respect to applications may be encountered.
- One example is given in the representation of scalar field intensities in which, by implementation of a closed range (ie. phase angle within complex plane), the lowest and highest intensities are numerically equivalent or similar.
- scalar fields are defined by values extending over either a bounded or unbounded region. An example of this may be field intensities or position, potentially extending over a range of 0 to infinity.
- inhibitory/excitory assignments must be distributed in such a fashion as to establish symmetry within the numerical process.
- Symmetry in this case refers to a uniform distribution of generated vectors for the correlation matrix [X] elements about the complex plane (ie. uniform distribution over the 2 ⁇ range). It is reiterated that maximum pattern storage capacity is attained when the distribution of both the stimulus and response vectors are uniform or symmetrical about the complex plane.
- One means of establishing symmetry would consist simply of creating a balanced distribution of inhibitory/excitory synaptic assignments.
- the sigmoid function or variation thereof may be employed to translate scalar quantities within the stimulus field into a near uniform probablistic distribution ranging between 0 and ⁇ .
- the sigmoid function is commonly used within control theory applications but, more significantly, has been determined empirically as the functional relationship by which receptor neurons respond to input stimulus as illustrated by Figure 3.
- the function for processing of stimulus input is given in the following general form:
- the synaptic or signal inhibitory/excitory classification is required to establish an open unbounded range for the stimulus/response signals and maintain symmetry of operation.
- the input stimulus values ranging between 0 and ⁇ as defined by equation 65 above
- the sign is reversed or made negative (0 to - ⁇ ) thus approaching the desired symmetry or uniform distribution extending over - ⁇ or + ⁇ .
- inhibitory/excitory assignment eliminates the inherent problems encountered for an implicitly closed range which folds onto itself, (ie. complex plane) and illustrates a means by which naturally occurring distributions (ie. Gaussian) may be mapped via the sigmoid function onto an optimal uniform probablistic distribution about the complex plane. Furthermore, this enhancement displays a similarity to the physical inhibitory/excitory synaptic structure of the neuron.
- This example of a hardware embodiment forms a conceptual basis for a possible future generation computer prototype using the process described herein.
- This device shall be referred to as a neurocomputer.
- the neurocomputer is intended to be general purpose in its operational characteristics as facilitated by a reprogramming utility.
- the programming methodology will be significantly different from conventional programming languages and techniques.
- the single chip microcomputer 70 is representative of a component currently available in the marketplace and contains the following features:
- the elements of the input stimulus field [S] and the output response field [R] are transmitted via their real and imaginary parts as binary values.
- the elements of the stimulus and the response fields may be transmitted either via the serial data links 80 and 82 or mapped onto the external memory devices (88 and 90).
- Figure 13 illustrates the transmission of one element of a stimulus pattern [S] via the serial data link 82.
- the elements of the stimulus pattern (s j ) are read in either via one of the four serial input links, or via the external memory device (88) providing memory mapped input.
- the element or elements of the response pattern may be read into the encoding unit 76 either via the serial links or via the external memory device 88. Note again that for each of the stimulus and response elements, a real and imaginary component derived from the phase and magnitude components of the complex element or vector are read into the encoding unit (76).
- the real and imaginary components of the corresponding element (x j ) for the correlation matrix [X] is read in from the external memory unit 86.
- the above operation superimposes the learned stimulus-response patterns onto the correlation matrix [X].
- the encoding operation is defined in further detail herein under the section entitled Theory. Note that the hardware configuration presented, and the operational characteristics of a general purpose processor, requires that the above matrix transformation be performed in a sequential manner.
- the encoded element or elements of the correlation matrix [X] are stored back into memory storage unit 86 which is believed to perform an analogous function to the axo-dendritic synaptic connections.
- the elements of the stimulus field (s* j ) are read in either via the four serial input links 82 or via the external memory device 88 for memory mapped input.
- a real and imaginary component derived from the phase and magnitude component of the complex element or vector (s* j ) are read into the decoding unit 74.
- the real and imaginary components of the corresponding complex or vector element within the correlation matrix [X] are read from the external memory unit 86.
- the generated response [R] may be output either via the serial output links 80 or via the external memory device 90 for memory mapped output. Note again that the response output signal is represented by a real and imaginary component as illustrated in equation 53 above.
- a third functional block defined within the processor block 78 is related to other functions or enhancements of the process.
- This functional block 78 has data access to both the serial input and output links 80 and 82 and the external memory 72.
- the embodiment of the ANS process within a general purpose processor provides a significant advantage in that a wide range of enhancements or variations of the basic process may be easily accommodated.
- a subset of possible enhancements consists of Hermittian or sigmoid preprocessing of the stimulus input field [S]* as described in the section pertaining to Enhancements of Invention, and all of the programmable functions listed under the heading Physical Embodiment of Device - Advanced Model ( Figure 13 and 14).
- FIG. 13 Several of the processing nodes, as depicted in Figure 13, may be connected together in any configuration to form an array of neural elements operating in an asycronous and parallel fashion.
- One possible parallel configuration is illustrated in Figure 14 and consists of 16 processing nodes.
- Each processing node 92 consists of the hardware configuration presented in Figure 13.
- the interconnection arrangement presented is generally referred to as a Boolean hypercube. In this arrangement each processing node accesses its own local memory and communicates with the four nearest nodes via high speed bidirectional serial communication.
- Each processing node may be configured internally to represent several distinct and separate neural analogues.
- the information stored within each of the 8000 synaptic input analogues, are represented by a complex or vector element within the correlation matrix [X].
- the memory requirements for each element of [X] for this example is 4 bytes.
- a 16-bit resolution for the real and imaginary elements within the correlation matrix [X] is assumed, accounting for the use of 4 bytes.
- stimulus-response patterns may be superimposed onto the same memory medium. As the number of patterns increases, the average magnitude of the deterministic error within the decoded output response increases in a non-linear fashion. As discussed previously it is assumed that for a non-ideal distribution, such as that produced via the encoding of arbitrary visual patterns, permits a capacity for encoded stimulus-response patterns that may be reduced over the ideal random case.
- the virtual information storage capacity corresponds to 16 Billion bytes.
- a second analogous system may be state space analysis as defined within control theory, however, this method is practicably limited to smaller systems in which a group of patterns may be most easily defined in a linear functional relationship.
- the state space analogy is limited within its functional domain and therefore potentially not applicable.
- the pattern matching technique involves a comparison of each element of the stimulus field against the corresponding element of each of the stored reference patterns.
- These patterns may represent any form of data either in non-compressed form or, if amenable, in a compressed form (i.e. converted to a frequency domain).
- a set of pattern variance terms would be generated as follows: ( ⁇ 1 2 , ⁇ 2 2 , ⁇ 3 2 , . . ⁇ p 2 ) and these values employed in some manner to determine or produce the output response.
- the assumed number of patterns encoded is 2000.
- Each comparison consists of evaluating the difference between an element in the input stimulus pattern and the corresponding element in a stored reference pattern and adjusting a correlation value (i.e. evaluation of a pattern variance or standard deviation value).
- a correlation value i.e. evaluation of a pattern variance or standard deviation value.
- the output response is derived directly from the decoding transformation, and is a function of all of the encoded patterns as defined by:
- the output response is issued directly as a vector quantity with phase angle representing the scalar response information and the magnitude statistically proportional to the degree of correspondence between stimulus and encoded patterns.
- a single transformation as described by equation 66 transforms one element within the stimulus field through the corresponding element for all of the encoded patterns (2000 in this example) superimposed in abstract form within matrix [X].
- the number of mechanical steps therefore required is therefore reduced by the proportional amount (i.e. in this case reduced by a factor of 1/2000).
- a single decoding operation (equation 19) simultaneously performs a virtual decoding operation on 2000 stimulus-response patterns.
- Each pattern represents one of 2000 stimulus-response patterns encoded into the abstract representation and superimposed onto correlation matrix [X].
- the hardware embodiment described herein therefore yields a capacity to store 16 Billion virtual elements (bytes) of information in the form of encoded stimulus-response patterns, and process input sensory stimulus through the above accumulated information at a rate of 32 Billion virtual operations / second. It is expected that a system of this type, receiving stimulus from various forms of sensory modalities (i.e. vision, sound, touch), may potentially provide a suitable information processing substrate to facilitate and evoke the emergent characteristics of a form of intelligence such as that exhibited by the most primitive insect neurobiological system.
- various forms of sensory modalities i.e. vision, sound, touch
- the described hardware embodiment is indicative of one possible example and may be expanded in any area including memory capacity, number of processing nodes, number of data links per node, rate of data link transmission, and processing rate per node. These increased capabilities may be used to construct devices emulating potentially higher levels of intelligence.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
- Supplying Of Containers To The Packaging Station (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
- Electrotherapy Devices (AREA)
Abstract
Description
Similarly the response field [R]' associated with the stimulus input field will be represented by:
or a = x + iy
where x = cos Θ , y = sin Θ
- Φp,k corresponds to the translated vector orientation for element k within pattern p for the response field[R].
- and p = 1 to P patterns are encoded
- N is the number of elements within the stimulus field
- elements rk, s*j and xj,k are vectors normally represented by complex numbers
- c is a normalization factor
The product of the above operation, the correlation matrix [X] is an N by 1 matrix where N is the number of elements in the stimulus field. This encoding operation (equation 35A) produces a matrix of the following form: The resultant effect of the above operation is that the dimension corresponding to time has been collapsed.
Information is stored within the ANS network via the superposition of data and the highest correlated response to an input stimulus is arrived at in a massively parallel fashion. It is expected that this inherently parallel technique will yield significant benefits over current heuristic methods in terms of both processing speed and information storage capacity.
local memory / node = 2 Megabytes
transmission rate / link = 20 Megabits/second
Processing speed / node = 20 Million Instructions / second
- (λ1, λ2, λ3, . . . λp)
- is the set of pattern correspondence values for patterns p = 1 to P
- (γ1, γ2, γ3, . . . γp)
- is the set of generated output response values for patterns p=1 to P
Estimate of Relative Error Within Response Field as a Function of Ratio P/N | |
Ratio of Encoded Patterns to # elements within Stimulus Field (P/N) | Error Within Response Field (%) |
0.05 | 2.5 |
0.1 | 3.4 |
0.15 | 4.2 |
0.2 | 4.7 |
0.25 | 5.2 |
0.3 | 5.6 |
0.35 | 6.0 |
0.4 | 6.3 |
0.45 | 6.6 |
0.5 | 6.9 |
Lilt of Phoneme Codes | |
CODE (base 6) | SOUND (capitalized in sample word) |
00 | mAke = m-A-k |
01 | bAt = b-AE-t |
02 | cAr = k-AH-r |
03 | dOg = d-AW-g |
04 | Bat = B-ae-t |
05 | CHeese = CH-ee-z |
10 | Dog = D-aw-g |
11 | bE = b-EE |
12 | bEt = b-EH-t |
13 | raFt = r-ae-F-t |
14 | Go = G-oh |
15 | Hive = H-i-v |
20 | tIme = t-I-m |
21 | sIt = s-IH-t |
22 | Jet = J-eh-t |
23 | Kill = K-ih-l |
24 | Love = L-uh-v |
25 | Map = M-ae-p |
30 | Nab = N-ae-b |
31 | gO = g-OH |
32 | gOO = g-OO |
33 | Pat = P-ae-t |
34 | Rat = R-ae-t |
35 | Sat = S-ae-t |
40 | SHe = SH-ee |
41 | Tap = T-ae-p |
42 | THin = TH-ih-n |
43 | THis = TZ-ih-s |
44 | wOrd = w-U-r-d |
45 | bUt = b-UH-t |
50 | Vat = V-ae-t |
51 | With = W-ih-th |
52 | WHich = WH-ih-ch |
53 | Yes = Y-eh-s |
54 | Zap = Z-ae-p |
55 | viSion = v-ih-ZH-eh-n |
Claims (8)
- An artificial neural device comprising:(a) means for storing and processing data;(b) means for decoding a response associated with a stimulus signal when said matrix of stimulus-response patterns encoded in said data storage means is stimulated by said stimulus signal, said artificial neural device being characterized in that it also includes(c) means for encoding data onto said data storage and processing means by translating sets of associated stimulus and response signals into normalized complex numbers represented by a vector phase angle and generating a matrix representing stimulus-response patterns for each set of associated stimulus and response signals, whereby each said stimulus-response pattern is superimposed onto said data storage means.
- An artificial neural device as claimed in Claim 1 wherein said data storing means comprise memory means and said device further includes:(a) means for inputting a stimulus signal into said device;(b) means for inputting and generating a response signal associated with a stimulus signal;(c) said processing means including:(i) data input means for communicating data from said stimulus signal to said processing means;(ii) data output means for communicating a response signal from said processing means in response to an associated stimulus signal;(iii) said encoding means reading the vector components of said stimulus signal from said data input means and reading the vector components of said response signal associated with said stimulus signal and generating a matrix of vector components representing stimulus-response patterns for each set of associated stimulus-response signals whereby each said stimulus-response pattern is superimposed onto said memory means;(iv) said decoding means reading the vector components of said stimulus signal from said data input means and reading the vector components of said matrix from said memory means and generating a response signal of vector components through said data output means; and(v) memory interface means for interfacing said processing means with said memory means.
- An artificial neural device as claimed in Claim 1 further including:(a) visual recording means for producing a stimulus signal representative of a visual image;(b) means for generating an auditory response signal associated with a stimulus signal representative of said visual image;(c) said means for encoding data onto said data storing and processing means including:(i) means for translating said visual stimulus signal into complex numbers represented by vector phase angles and storing said translated visual stimulus signals onto said data storage means;(ii) means for translating said auditory response signals into complex numbers represented by vector phase angles and storing said translated auditory response signals onto said data storage means;(iii) means for generating and storing a matrix representing stimulus-response patterns for each set of associated visual stimulus and auditory response signals and storing each said stimulus-response patterns onto said data storage means;(d) said means for decoding a response associated with a stimulus signal includes means for generating an auditory response signal when said matrix of stimulus-response patterns is stimulated by said visual stimulus signal associated with said auditory response signal;
- An artificial neural device as claimed in claim 1 further comprising:(a) local memory storage means for storing data mapped onto said memory storage means;(b) micro-processing means including:(i) processor means;(ii) a plurality of data input means for receiving stimulus signals;(iii) a plurality of data output means for transmitting response signals;(iv) memory interface means for interfacing said microprocessor means with said local memory storage means;(c) said processing means including:(i) means for encoding data onto said processing means and local memory storage means by translating sets of associated stimulus and response signals, received by said data input means or mapped onto said local memory storage means, into complex numbers represented by a vector phase angle and a magnitude, and generating a matrix representing stimulus-response patterns for each set of associated stimulus and responses signals, whereby each said stimulus-response pattern is superimposed onto said local memory storage means;(ii) means for decoding a response associated with a stimulus signal received by said data input means or mapped onto said local memory means when said matrix of said stimulus-response patterns encoded on said local memory storage means is stimulated by said stimulus signal so as to generate a response signal associated with said stimulus signal through said data output means or to said local memory storage means.
- A plurality of artificial neural devices as claimed in Claim 4 whereby each said artificial neural device accesses its own local memory storage means and communicates with a plurality of other artificial neural devices as claimed in Claim 4 by means of bidirectional communication means.
- A method of enhancing the storage and computing capacity of artificial neural devices comprising the steps of:(a) decoding a response signal associated with a stimulus signal by applying said stimulus signal to said stimulus-response patterns encoded in said data storage means so as to generate said response signal associated with said stimulus signal, said method being further characterized by the following steps :(b) encoding data by translating sets of associated stimulus and response signals into normilized complex numbers represented by a vector phase angle and magnitude;(c) generating a matrix representing stimulus-response patterns for each set of associated stimulus-response signals whereby each said stimulus-response is superimposed onto data storage and processing means.
- A method as claimed in Claim 6 including the steps of:(a) producing a stimulus signal;(b) generating a response signal associated with said stimulus signal;(c) said encoding step comprising:(i) translating said stimulus signal by processing means into complex numbers represented by vector phase angle and magnitude;(ii) storing said translated stimulus signal into storage means;(iii) translating said response signals associated with said stimulus signal by processing means into complex numbers represented by vector phase angle and magnitude;(iv) storing said translated response signal onto said storage means.
- A method as claimed in Claim 7 further including the steps of :(a) producing a visual stimulus signal through visual recording means;(b) generating an auditory response signal associated with said visual stimulus signal;
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CA575483 | 1988-08-23 | ||
CA575483 | 1988-08-23 |
Publications (3)
Publication Number | Publication Date |
---|---|
EP0357212A2 EP0357212A2 (en) | 1990-03-07 |
EP0357212A3 EP0357212A3 (en) | 1991-07-03 |
EP0357212B1 true EP0357212B1 (en) | 1998-04-29 |
Family
ID=4138600
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
EP89307440A Expired - Lifetime EP0357212B1 (en) | 1988-08-23 | 1989-07-21 | Artificial neural device and corresponding method |
Country Status (4)
Country | Link |
---|---|
EP (1) | EP0357212B1 (en) |
JP (1) | JPH02236768A (en) |
AT (1) | ATE165680T1 (en) |
DE (1) | DE68928656D1 (en) |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
ATE147873T1 (en) * | 1990-04-26 | 1997-02-15 | John Sutherland | ARTIFICIAL NEURONAL EQUIPMENT |
CA2040903C (en) * | 1991-04-22 | 2003-10-07 | John G. Sutherland | Neural networks |
WO2002054951A2 (en) * | 2001-01-12 | 2002-07-18 | The Government Of The United States Of America As Represented By The Secretary, Department Of Health And Human Services | Decoding algorithm for neuronal responses |
CA2883091C (en) * | 2011-08-25 | 2020-02-25 | Cornell University | Retinal encoder for machine vision |
US11849286B1 (en) | 2021-10-25 | 2023-12-19 | Chromatic Inc. | Ear-worn device configured for over-the-counter and prescription use |
US12075215B2 (en) | 2022-01-14 | 2024-08-27 | Chromatic Inc. | Method, apparatus and system for neural network hearing aid |
US11950056B2 (en) | 2022-01-14 | 2024-04-02 | Chromatic Inc. | Method, apparatus and system for neural network hearing aid |
US11832061B2 (en) | 2022-01-14 | 2023-11-28 | Chromatic Inc. | Method, apparatus and system for neural network hearing aid |
US20230306982A1 (en) | 2022-01-14 | 2023-09-28 | Chromatic Inc. | System and method for enhancing speech of target speaker from audio signal in an ear-worn device using voice signatures |
US11818547B2 (en) * | 2022-01-14 | 2023-11-14 | Chromatic Inc. | Method, apparatus and system for neural network hearing aid |
EP4333464A1 (en) | 2022-08-09 | 2024-03-06 | Chromatic Inc. | Hearing loss amplification that amplifies speech and noise subsignals differently |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4366551A (en) * | 1977-06-24 | 1982-12-28 | Holtz Klaus E | Associative memory search system |
JPS58129684A (en) * | 1982-01-29 | 1983-08-02 | Toshiba Corp | Pattern recognizing device |
US4523331A (en) * | 1982-09-27 | 1985-06-11 | Asija Satya P | Automated image input, storage and output system |
-
1989
- 1989-07-21 AT AT89307440T patent/ATE165680T1/en active
- 1989-07-21 DE DE68928656T patent/DE68928656D1/en not_active Expired - Lifetime
- 1989-07-21 EP EP89307440A patent/EP0357212B1/en not_active Expired - Lifetime
- 1989-08-23 JP JP1217171A patent/JPH02236768A/en active Pending
Also Published As
Publication number | Publication date |
---|---|
EP0357212A3 (en) | 1991-07-03 |
ATE165680T1 (en) | 1998-05-15 |
EP0357212A2 (en) | 1990-03-07 |
DE68928656D1 (en) | 1998-06-04 |
JPH02236768A (en) | 1990-09-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US5214745A (en) | Artificial neural device utilizing phase orientation in the complex number domain to encode and decode stimulus response patterns | |
Pandya et al. | Pattern recognition with neural networks in C++ | |
Prieto et al. | Neural networks: An overview of early research, current frameworks and new challenges | |
Kohonen | Adaptive, associative, and self-organizing functions in neural computing | |
Anderson et al. | Artificial neural networks technology | |
Robinson et al. | The utility driven dynamic error propagation network | |
Yegnanarayana | Artificial neural networks for pattern recognition | |
EP0526504B1 (en) | Artificial neural device | |
US5255348A (en) | Neural network for learning, recognition and recall of pattern sequences | |
US6038338A (en) | Hybrid neural network for pattern recognition | |
EP0357212B1 (en) | Artificial neural device and corresponding method | |
Yarushev et al. | Time series analysis based on modular architectures of neural networks | |
Card et al. | Competitive learning algorithms and neurocomputer architecture | |
Basegmez | The next generation neural networks: Deep learning and spiking neural networks | |
Alspector | Neural-style microsystems that learn | |
Peterson | Neural networks and high energy physics | |
Zhao et al. | Efficient learning of NN-MLP based on individual evolutionary algorithm | |
De Jongh, PJ & De Wet | An introduction to neural networks | |
Poon et al. | Hebbian learning in parallel and modular memories | |
Yoon | Artificial neural network technology | |
Stomski et al. | Selection of a neural network system for visual inspection | |
Yang et al. | A Computational Framework of Cortical Microcircuits Approximates Sign-concordant Random Backpropagation | |
Myers | Neural networks in communications | |
Hartono | Ensemble of linear experts as an interpretable piecewise-linear classifier | |
KR930007744B1 (en) | Optical neural network having reduced variable interconnection |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
AK | Designated contracting states |
Kind code of ref document: A2 Designated state(s): AT BE CH DE ES FR GB GR IT LI LU NL SE |
|
PUAL | Search report despatched |
Free format text: ORIGINAL CODE: 0009013 |
|
AK | Designated contracting states |
Kind code of ref document: A3 Designated state(s): AT BE CH DE ES FR GB GR IT LI LU NL SE |
|
17P | Request for examination filed |
Effective date: 19911213 |
|
17Q | First examination report despatched |
Effective date: 19950504 |
|
GRAG | Despatch of communication of intention to grant |
Free format text: ORIGINAL CODE: EPIDOS AGRA |
|
GRAG | Despatch of communication of intention to grant |
Free format text: ORIGINAL CODE: EPIDOS AGRA |
|
GRAH | Despatch of communication of intention to grant a patent |
Free format text: ORIGINAL CODE: EPIDOS IGRA |
|
GRAH | Despatch of communication of intention to grant a patent |
Free format text: ORIGINAL CODE: EPIDOS IGRA |
|
GRAA | (expected) grant |
Free format text: ORIGINAL CODE: 0009210 |
|
AK | Designated contracting states |
Kind code of ref document: B1 Designated state(s): AT BE CH DE ES FR GB GR IT LI LU NL SE |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: IT Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT;WARNING: LAPSES OF ITALIAN PATENTS WITH EFFECTIVE DATE BEFORE 2007 MAY HAVE OCCURRED AT ANY TIME BEFORE 2007. THE CORRECT EFFECTIVE DATE MAY BE DIFFERENT FROM THE ONE RECORDED. Effective date: 19980429 Ref country code: FR Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980429 Ref country code: ES Free format text: THE PATENT HAS BEEN ANNULLED BY A DECISION OF A NATIONAL AUTHORITY Effective date: 19980429 Ref country code: CH Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980429 Ref country code: BE Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980429 Ref country code: AT Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980429 Ref country code: GR Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980429 Ref country code: LI Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980429 Ref country code: NL Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980429 |
|
REF | Corresponds to: |
Ref document number: 165680 Country of ref document: AT Date of ref document: 19980515 Kind code of ref document: T |
|
REG | Reference to a national code |
Ref country code: CH Ref legal event code: EP |
|
REF | Corresponds to: |
Ref document number: 68928656 Country of ref document: DE Date of ref document: 19980604 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: LU Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 19980721 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: SE Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980729 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: DE Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 19980730 |
|
EN | Fr: translation not filed | ||
NLV1 | Nl: lapsed or annulled due to failure to fulfill the requirements of art. 29p and 29m of the patents act | ||
REG | Reference to a national code |
Ref country code: CH Ref legal event code: PL |
|
PLBE | No opposition filed within time limit |
Free format text: ORIGINAL CODE: 0009261 |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: NO OPPOSITION FILED WITHIN TIME LIMIT |
|
26N | No opposition filed | ||
REG | Reference to a national code |
Ref country code: GB Ref legal event code: IF02 |
|
PGFP | Annual fee paid to national office [announced via postgrant information from national office to epo] |
Ref country code: GB Payment date: 20080714 Year of fee payment: 20 |
|
REG | Reference to a national code |
Ref country code: GB Ref legal event code: PE20 Expiry date: 20090720 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: GB Free format text: LAPSE BECAUSE OF EXPIRATION OF PROTECTION Effective date: 20090720 |